Sorry - It was an abstract at the HBM conference in Sendai in 2002.
http://www.academicpress.com/journals/hbm2002/14693.html
It is currently being written up as a paper.
thanks
Tim
On Thu, 11 Jul 2002, Goldfine, Andrew (NINDS) wrote:
> I couldn't find Jenkinson,Beckmann et al. HBM 2002. Is it published yet?
>
> -----Original Message-----
> From: Tim Behrens [mailto:[log in to unmask]]
> Sent: Thursday, July 11, 2002 12:35 PM
> To: [log in to unmask]
> Subject: Re: [FSL] Event-related design
>
> Hi Joe and Andrew
>
> When doing a multisession multisubject analysis, there are various
> variances you might want to consider.
>
> 1 - First level variance: This is the variance in the COPEs from an
> individual analysis (VarCOPE) caused by factors such as MR noise.
>
> 2 - Between Session Variance - Subjects will behave slightly differently
> in different sessions, leading to a variance in the mean response in a
> single subject between sessions.
>
> 3 - Group (between subject) Variance - This variance models the inherent
> variability of different subjects from the same population.
>
> 4 - Between Group Variance
>
>
> When all these variances combine to give the final Variance that is
> estimated by Feat/SPM at the group level, this is known as a Mixed effects
> Variance.
>
>
> The two approaches laid out in Joe's email represent two different models
> for the analysis.
>
> 1) The Feat/SPM approach:
>
> Combine all sessions from all subjects together implicitly assumes that
> the between Session Variance and the between subject variance are the
> same, so that it doesn't matter that we choose to estimate them as one.
>
> This is clearly not a very good assumption to make - However it will
> increase the degrees of freedom to estimate the Mixed Effects Variance.
>
> 2) Joe's suggested Method:
>
> Perform a fixed effects analysis betweend different session of the same
> subject, and a random effects analysis on the resulting COPEs.
>
> This method does not suffer from the assumption made above and thus, in
> a sense is more valid. By combining the sessions in a fixed effects
> analysis, you are effectively reducing the variance in the estimate of
> your associated mean COPE. Unfortunately, neither Feat nor SPM carry
> this information through to the next level. Therefore the only advantage
> you have gained, is a small reduction in your mixed effects variance, at
> the expense of a reduction in the degrees of freedom available to estimate
> the variance at the highest level.
>
>
> It turns out that the most efficient estimator for the correct model needs
> to carry variance information from one level to the next.
> See
>
> Jenkinson,Beckmann et al. HBM 2002
>
> which will hopefully be available in the next release of Feat.
>
>
> Hope this all makes sense.
>
> Thanks
>
> Christian and Tim
>
>
> On Thu, 11 Jul 2002, Joe Devlin wrote:
>
> > Hi Andrew,
> >
> > >If each run is analysed separately, do I combine them as if they are
> > >separate subjects doing the experiment or is there a different method in
> > >this case?
> >
> > Well, you could either. The first method -- combining them as if they
> were
> > separate subjects, is what SPM does. You can do the same in Feat by
> > choosing a group stats analysis and entering each session into the whole.
> >
> > Alternatively you could manually create mean contrast of parameter
> estimate
> > (COPE) images per subject and then use Medx to compute a group random
> > effects analysis form the subjects' mean PE files. I've attached a script
> > I've used to do this method, again in ksh. This one isn't commented as
> > much and could be made more elegant but it follows the same pattern.
> >
> > Once you have the mean images per subject (in standard space), you need to
> > load them into a group in medx and choose Toolbox -> Functional -> Group
> > Statistics. Then set the operation to single group t-test and run it on
> > your grouped COPE images. I'm not sure how to script this part from
> within
> > the rfx shell script so I do it manually. Afterwards, if you save the
> > z-image, you can run a second script called thresh_render (attached) to
> > make the various images that you normally get from Feat.
> >
> > This isn't completely elegant but it does compute a type of repeated
> > measures RFX analysis which is more correct that simply treating the
> > various runs as different subjects.
> >
> > Perhaps one of the more statistically-minded members of the list will
> > comment on that.
> >
> > Cheers,
> > Joe
> >
> >
> >
> >
> >
> > Joe
>
> --
> ----------------------------------------------------------------------------
> ---
> Tim Behrens
> Centre for Functional MRI of the Brain
> The John Radcliffe Hospital
> Headley Way Oxford OX3 9DU
> Oxford University
> Work 01865 222782
> Mobile 07980 884537
> ----------------------------------------------------------------------------
> ---
>
--
-------------------------------------------------------------------------------
Tim Behrens
Centre for Functional MRI of the Brain
The John Radcliffe Hospital
Headley Way Oxford OX3 9DU
Oxford University
Work 01865 222782
Mobile 07980 884537
-------------------------------------------------------------------------------
|